Clinical AI Safety / Learning by doing
The judgement
behind the answer.
I’m building an interactive curriculum that helps clinicians question AI outputs, inspect the evidence and work through the safety decisions that follow.
The question
What does safe use look like
in the moment?
A clinician needs more than a description of how a model works. They need practice noticing a changed fact, finding the original evidence and deciding what still needs checking. The curriculum puts those decisions inside fictional cases.
01 / Try the teaching
Read it. Question it. Check it.
A short excerpt from Module 2
The verification desk
0 / 3 statements explored
Same record.
Different meaning?
Select a detail. Compare the source with the simulated summary, then decide whether its meaning has changed.
Fictional teaching material. These excerpts are authored examples, not live model responses. Your choices stay in this page.
01 / Original source
She has a documented penicillin allergy (rash), so she was treated with doxycycline.
02 / Simulated summary
“No known drug allergies.”
02 / The wider curriculum
From everyday use
to a defensible decision.
A shared foundation for clinicians, followed by specialist safety practice. Open a module to see its purpose. Developed lessons and planned material are shown separately.
- 14
- modules in the plan
- 05
- developed lesson modules
A clinician’s foundation
Use · Verify · Disclose · Escalate
M2What AI actually is: a clinician's working modelDeveloped lesson
Rules vs machine learning vs large language models; training data; why outputs are probabilistic; retrieval and agents; why the same question can get different answers.
6 units · 5 case discussions · 14 assessments
M3How AI fails: a catalogue of failure modesDeveloped lesson
Name and recognise each AI failure mode, and map it to hazard-log cause categories so it can be reported and controlled.
6 units · 5 case discussions · 15 assessments
M4AI at the point of care: documentation and decision supportDeveloped lesson
What the evidence says about ambient scribes and record summarisers, and about triage, imaging and decision-support tools — what they get wrong and what to check before you act or sign.
6 units · 5 case discussions · 15 assessments
M5Using AI responsibly: disclose, protect, escalateOutline
Explain AI involvement and its limits to patients; protect patient data when using AI tools; recognise and report an AI-related incident; speak up.
Specialist safety practice
Evidence · Controls · Assurance · Monitoring
M1Foundations of clinical safety for health IT and AIDeveloped lesson
The CRM process, DCB0129/0160, roles, the risk matrix, the hazard log and the safety case, with first AI examples.
6 units · 5 case discussions · 15 assessments
M6Regulatory status, procurement and supplier evidenceDeveloped lesson
Is it a medical device? Intended purpose, MHRA AI-as-a-medical-device status, UKCA/CE; what to demand from a supplier (DTAC, NICE Evidence Standards Framework, contract clauses, change notification) before assurance starts.
6 units · 5 case discussions · 14 assessments
M7From evidence to safety requirements: appraising AI and arguing its safety caseOutline
Appraise AI claims (metrics, benchmark vs prospective evidence, leakage, calibration, external validation) and turn what you find into safety requirements and an AMLAS/DCB0129 safety case.
M8Data protection, cybersecurity and equity for AIOutline
DPIA and lawful basis for AI, confidentiality of prompts and transcripts, data residency; threat-modelling AI including prompt injection; subgroup performance and the Equality Act.
M9Deploying predictive and diagnostic AI (DCB0160)Outline
Local validation, thresholds and alert burden, uncertainty gating, human-oversight design and fallback for risk scores, imaging and triage models.
M10Deploying generative and agentic AI (DCB0160)Outline
Guardrails, fail-closed escalation floors, deterministic safety layers, grounding and citation checks, stopping conditions and human-on-the-loop oversight for LLM tools and agents.
M14Designing for safety and sharing the hazard: interface, human factors and organisational learningOutline
How the interface itself becomes a clinical safety control, how oversight is designed rather than assumed, how hazards are shared between manufacturer and deploying organisation — and how an organisation learns from what its AI does.
M11AI in live use: monitoring, incidents, change and decommissioningOutline
Post-deployment surveillance, drift detection, learning from incidents, model and EPR updates under change control, staged authorisation and safe decommissioning.
M12Governance, accountability and liabilityOutline
The AI register, the accountable executive, approval boundaries, who carries liability — and why the CSO assures a governed decision rather than owning it alone.
M13Capstone: the assurance labOutline
Take a real open-source clinical AI tool, red-team it with open evaluation harnesses, and write its hazard log and clinical safety case report for peer review.
The design decisions
Make the reasoning inspectable.
- Start with a decision
- Give the learner something to judge before revealing the explanation. Make the difference between a wording change and a factual change concrete.
- Keep the source close
- Let people compare the statement with the evidence that supports it. Research references, uncertainty and limitations belong beside the teaching.
- Carry the lesson into practice
- The wider project includes a verification drill, hazard-log workbench and safety-case builder, connecting the exercise to an artifact a learner can inspect.
Development and evidence notes
This portfolio exercise adapts three comparisons from “Same question, different answer”. The original records and summaries are fictional. Research statistics from the longer exercise are not reproduced here.
The curriculum uses AI-assisted research and drafting, with source-linked teaching material. Source preservation does not establish independent evidence review. Clinical review and evaluation of learning outcomes remain to be completed; this is not an accredited course.
From teaching to a working tool